As the agent is informed of the changes to the external world -
through its own actions or through the actions of other agents - it
can gain and lose LCW; these changes must be recorded in . We
assume here, and throughout, the absence of hidden exogenous
events that invalidate XII 's information. In other words, we assume
that the rate of change in the world is slower than the rate at which
XII plans and executes. This is the standard assumption of correct
information made by most planners.
When XII executes an operator which ensures that contains all
instances of
that are true in the world, XII adds a sentence
LCW(
) to
. For example, XII is given an axiom stating
that each file has a unique word count. Thus, executing the UNIX
command wc paper.tex adds the sentence
LCW(word.count(paper.tex,
))
to
as well as adding the actual length (e.g., word.count(paper.tex,42)) to
. Since the ls command has a
universally quantified effect, executing ls /tex yields LCW(parent.dir(
,/tex)).
It would be cumbersome if the author of each operator were forced to list its LCW effects. In fact, this is unnecessary. XII automatically elaborates operator schemata with LCW effects. Even in the worst case, this compilation process takes time linear in the length of the operator schemata and the number of unique-value axioms [8].
Observational effects (e.g., those of ls) can only
create LCW, but causal effects can both create and destroy
LCW. For example, deleting all files in
/tex provides complete information on the contents of the
directory regardless of what the agent knew previously. Compressing a
file in /tex, on the other hand, invalidates previously obtained
LCW of the lengths of all files in that directory.
The theory behind LCW is complex; [3] defines LCW formally, explains the connection to circumscription, and presents a set of tractable update rules for the case of conjunctive LCW sentences. In this paper, we show how to incorporate conjunctive LCW into a least commitment planner and argue that this addresses the challenges described in section 1: satisfying universally quantified goals and avoiding redundant sensing.